Anthropic has signed the EU AI Act Code of Practice, announcing that all new released Claude models will be embedded with content identifiers globally starting August 2026. This requirement applies not only to the EU market but also covers all Claude product lines including API, Claude, Claude Code, Claude Cowork, and Claude Tag. Existing models will have a legal transition period, but Anthropic states it is already working on retroactive adaptation of the watermark feature.
Anthropic adopts two identification mechanisms: text output will carry an invisible watermark that does not affect the meaning or readability of the content, and the watermark will remain even after copy-pasting and "may persist to some extent after editing." File formats such as images will add digital signature traceability metadata based on the C2PA open standard, which can indicate that the file has been processed by Claude and reveal subsequent tampering. Text watermarks will also take effect through cloud partners such as AWS, Google Cloud, and Microsoft Foundry, although these platforms may not support signature metadata. Anthropic plans to release verification tools later for users and third parties to detect these identifiers.
Watermark detection has inherent limitations
Anthropic acknowledges clear limitations in watermark technology: detecting a watermark does not mean the content was entirely generated by Claude, as users often use Claude for proofreading, translation, or summarization, and the output content may carry a watermark but the core idea comes from humans. Similarly, the absence of detected watermarks cannot exclude the involvement of AI—models may have generated the text before the watermark feature was launched, the text may have been heavily edited or translated, paragraphs may be too short to detect reliably, or metadata may be stripped during format conversion and screenshots. Whether the watermark can survive after editing, reformatting, and translation will be the key test of its true value.
AI text detection has significant implications in socially sensitive areas such as education. Research shows that over-reliance on AI tools can weaken critical thinking and writing skills, and there have been cases of scammers using AI to register fake student identities to obtain scholarships. However, unreliable detectors are also dangerous, potentially leading to erroneous accusations of students cheating. If Anthropic's watermark solution is more reliable than existing third-party detection tools, it could offer a new way out of this dilemma.
Industry Landscape: Diverging Attitudes Among Companies
Anthropic is not the only company exploring watermark technology. Google DeepMind has open-sourced the SynthID watermark system and integrated it into Gemini models, achieving watermarking without quality loss by fine-tuning probability values during token prediction, supporting multiple languages but facing similar challenges in detection after editing. In contrast, OpenAI has an accurate text detector with a 99.9% accuracy rate but has delayed its release for two years, citing reasons including users being able to easily bypass it through translation or rewriting, potential stigmatization of specific groups, and public detectors possibly harming its own commercial interests.
Anthropic's decision may also bring commercial risks. Claude is particularly popular among students in knowledge work fields, and more reliable detection capabilities may reduce the appeal of this user group. However, as AI-generated content becomes widespread and deepfake risks increase, content tracing has moved from an option to a necessity. Vendors who establish a trusted identification system first will gain a first-mover advantage in an environment of tightening regulation.
